Download PDF - Ivie
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
ad<br />
serie<br />
WP-AD 2011-14<br />
On downside risk predictability through<br />
liquidity and trading activity:<br />
a quantile regression approach<br />
Antonio Rubia and Lidia Sanchis-Marco<br />
Working papers<br />
g papers<br />
Working papers
Los documentos de trabajo del <strong>Ivie</strong> ofrecen un avance de los resultados de las<br />
investigaciones económicas en curso, con objeto de generar un proceso de<br />
discusión previo a su remisión a las revistas científicas. Al publicar este<br />
documento de trabajo, el <strong>Ivie</strong> no asume responsabilidad sobre su contenido.<br />
<strong>Ivie</strong> working papers offer in advance the results of economic research under way<br />
in order to encourage a discussion process before sending them to scientific<br />
journals for their final publication. <strong>Ivie</strong>’s decision to publish this working paper<br />
does not imply any responsibility for its content.<br />
La Serie AD es continuadora de la labor iniciada por el Departamento de<br />
Fundamentos de Análisis Económico de la Universidad de Alicante en su<br />
colección “A DISCUSIÓN” y difunde trabajos de marcado contenido teórico.<br />
Esta serie es coordinada por Carmen Herrero.<br />
The AD series, coordinated by Carmen Herrero, is a continuation of the work<br />
initiated by the Department of Economic Analysis of the Universidad de<br />
Alicante in its collection “A DISCUSIÓN”, providing and distributing papers<br />
marked by their theoretical content.<br />
Todos los documentos de trabajo están disponibles de forma gratuita en la web<br />
del <strong>Ivie</strong> http://www.ivie.es, así como las instrucciones para los autores que<br />
desean publicar en nuestras series.<br />
Working papers can be downloaded free of charge from the <strong>Ivie</strong> website<br />
http://www.ivie.es, as well as the instructions for authors who are interested in<br />
publishing in our series.<br />
Edita / Published by: Instituto Valenciano de Investigaciones Económicas, S.A.<br />
Depósito Legal / Legal Deposit no.: V-2189-2011<br />
Impreso en España (junio 2011) / Printed in Spain (June 2011)
WP-AD 2011-14<br />
On downside risk predictability through<br />
liquidity and trading activity:<br />
a quantile regression approach<br />
Antonio Rubia and Lidia Sanchis-Marco *<br />
Abstract<br />
Most downside risk models implicitly assume that returns are a sufficient statistic with which to<br />
forecast the daily conditional distribution of a portfolio. In this paper, we address this question<br />
empirically and analyze if the variables that proxy for market liquidity and trading conditions<br />
convey valid information to forecast the quantiles of the conditional distribution of several<br />
representative market portfolios. Using quantile regression techniques, we report evidence of<br />
predictability that can be exploited to improve Value at Risk forecasts. Including trading- and<br />
spread-related variables improves considerably the forecasting performance.<br />
Keywords: Value at Risk, Basel, Liquidity, Trading Activity.<br />
* We would like to thank David Abad, Jos van Bommel, Lucía Cuadro, Christian Riis Flor, Angel León,<br />
Antonio Machado, Simone Manganelli, Manuel Moreno, Alfonso Novales, Alberto J. Plazzi, Gonzalo<br />
Rubio, Allan Timmerman, Ross Valkanov, participants in XVII Finance Forum (Madrid, Spain), EFMA<br />
2010 Meeting (Aarhus, Denmark), EAA 2010 Meeting (Istanbul, Turkey), 2010 ASEPELT Meeting<br />
(Alicante, Spain) and seminar attendants at University of Valencia for their valuable comments and<br />
suggestions to previous versions of this paper. Any error is our own responsibility. Financial support from<br />
the Spanish Department of Science and Innovation through ECO2008-02599 project and <strong>Ivie</strong> is gratefully<br />
acknowledged. A. Rubia: University of Alicante. L. Sanchis-Marco: University of Castilla-La Mancha.<br />
Corresponding author: lidia.sanchis@uclm.es.
1 Introduction<br />
Implementing risk control and monitoring systems requires quantitative procedures<br />
to capture the level of underlying uncertainty and make accurate predictions. The<br />
Basel Committee on Banking Supervision (BCBS) has contributed greatly to the<br />
popularization of certain international standards, known as Basel I and II Accords,<br />
in the …nancial services industry. This regulatory setting entitles eligible …nancial<br />
institutions to use internal models based on the Value-at-Risk (VaR) measure for<br />
meeting market risk capital requirements. It is agreed that, without the e¤orts<br />
made to comply with the BCBS standards, the …nancial industry would likely be<br />
facing a deeper crisis. The depth of the economic turmoil has shown the necessity<br />
for new standards to achieve greater banking sector resilience and the need to<br />
improve the existing procedures for quantifying market risk. 1 The present paper<br />
is motivated by this concern.<br />
The existing literature has suggested a number of procedures for forecasting<br />
downside risk, mainly VaR, which largely di¤er in their degree of sophistication:<br />
From the simple EWMA approach popularized by RiskMetrics to the more advanced<br />
probabilistic settings based on the Extreme Value Theory; see Manganelli<br />
and Engle (2004) and McNeil et al. (2005) for a review. Remarkably, a number<br />
of empirical studies have revealed that most of these methods do not seem to perform<br />
successfully in practice under standard backtesting techniques (e.g., Kuester<br />
et al., 2006), which underlines the practical complexity that lies behind the simple<br />
notion of VaR. Why is accurate VaR forecasting so elusive? Whereas most<br />
of the previous literature has attempted to address this question on the grounds<br />
of model misspeci…cation, in this paper we adopt an alternative view within the<br />
framework of model risk and analyze the important role played by the set of conditioning<br />
information. In spite of the large methodological di¤erences, the existing<br />
methodologies to model market risk share a common characteristic: They all rely<br />
exclusively on the information conveyed by historical returns. Naturally, this may<br />
turn out to be unnecessarily restrictive for practical purposes, since the conditional<br />
loss function of a portfolio may exhibit non-trivial links with the state variables<br />
that characterize the market environment and trading conditions and which may<br />
help forecasting bursts in volatility and illiquidity, particularly, in times of stress. 2<br />
1 The BCBS is currently carrying out a reform program that aims to enhance Basel II towards<br />
a new regulatory setting on the basis of higher supervisory standards (Basel III), and which is expected<br />
to be completed by the end of 2010. The BCBS has issued several consultative documents<br />
in the interim outlining the main features of the new regulatory setting. It is interesting to note<br />
that, despite its limitations, the VaR paradigm is still a pillar in the internal model approach for<br />
market risk adopted in Basel III, although further re…nements, such as stressed VaR and stress<br />
tests, will also be required to provide a complementary risk perspective.<br />
2 The implicit belief that returns subsume all the relevant information to forecast downside<br />
risk may be originated in a conservative interpretation of the E¢ cient Market Hypothesis. This<br />
1<br />
3
In this paper, we address empirically the premise that certain state variables<br />
that are observable in the market trading process exhibit predictive power to forecast<br />
the tail of the conditional loss distribution and, consequently, are useful for<br />
risk management. Although predictability is not necessarily limited to the set of<br />
variables we analyze, our main focus is on bid-ask spread and volume-related measures.<br />
Our study is motivated by previous …ndings and theoretical considerations<br />
in the asset pricing and market microstructure literature which have underlined<br />
the link between returns and market liquidity, activity, and private information<br />
arrivals. Like returns, liquidity- and volume-related variables are available on the<br />
trading-basis and are highly sensitive to information ‡ow. Like volatility, these<br />
variables are believed to re‡ect collective expectations, environmental conditions<br />
and market sentiments which have a major in‡uence on investor decisions. In<br />
contrast to returns and volatility, however, trade-related variables seem to have<br />
been ignored in the literature devoted to downside risk modelling, even though<br />
there exists previous evidence that supports the predictive power of volume and<br />
liquidity variables on volatility (Suominen, 2001; Bollerslev and Melvin, 1994).<br />
The main aim of this paper, therefore, is to analyze empirically whether downside<br />
risk forecasts can be enhanced by using this information or if, on the contrary,<br />
the predictability of the conditional loss distribution is limited to past returns<br />
and their volatility, as implicitly assumed in most of the empirical models used in<br />
practice.<br />
More speci…cally, we study the performance of several liquidity- and volumerelated<br />
predictors in day-ahead VaR forecasting at di¤erent quantiles using daily<br />
log returns from volume- and value-weighted market portfolios, book-to-market<br />
(B/M), and size-sorted portfolios from the US Stock Exchange in the period<br />
01/1988 through 12/2002. 3 As in the literature concerned with return predictability,<br />
the simplest way to appraise forecastability is to use simple least-squares analysis<br />
in predictive linear regressions; see, for instance, Cochrane (2005). However,<br />
conditional quantiles are unobservable and have to be estimated or, at best, modelled<br />
as a latent process, which makes such an approach infeasible. Fortunately, the<br />
Quantile Regression theory (Koenker and Bassett, 1978) allows us to analyze formally<br />
predictability in the conditional percentiles without departing signi…cantly<br />
from the intuitive spirit that characterizes predictive regressions. Using quantile<br />
regression we can directly model the tail of the conditional distribution of returns<br />
through a functional form that relates the VaR time-varying dynamics to its own<br />
forbids the systematic predictability of returns on the basis of the available information, i.e.,<br />
posits an orthogonal condition on the …rst-order conditional moment. However, it remains silent<br />
about higher-order moments, such as conditional volatility, or other distributional features of<br />
returns, such as conditional percentiles. Furthermore, …nancial markets largely depart in practice<br />
from the complete-market and symmetric-information hypotheses that underlie a number of<br />
theoretical models in the asset pricing literature. In the presence of asymmetric or imperfect<br />
information and other frictions, even the conditional mean of returns may be predictable.<br />
3 Three main reasons prompted us to consider this speci…c sample: i) market-portfolio data<br />
allow us to eliminate the idiosyncratic noise that may a¤ect the main conclusions on individual<br />
stocks, ii) the sample period is particularly interesting for risk management as it includes a stress<br />
scenario originating in the burst of the technological bubble in 2000, and iii) we can analyze the<br />
aggregate measures of liquidity and volume that have been used previously in several studies;<br />
see Chordia et al. (2001). We thank Prof. A. Subrahmanyam for making these data available.<br />
2<br />
4
past as well as to lagged predictors without making an explicit assumption on<br />
the distribution of returns, building on the so-called CAViaR model setting in<br />
Engle and Manganelli (2004). A restricted version of this model, which considers<br />
returns-related information solely, can be taken as a natural benchmark to address<br />
statistically the incremental value of liquidity- and volume-related variables.<br />
Even more importantly, these models can be used to construct VaR forecasts,<br />
which allows us to resort to standard backtesting and statistical techniques (e.g.,<br />
Christo¤ersen, 1998; Piazza et al., 2009) to analyze the actual out-of-sample performance.<br />
Our main empirical conclusions largely support the suitability of the<br />
liquidity- and volume-related variables in forecasting daily VaR.<br />
This paper belongs to the stream of literature that has focused on quantile<br />
regressions for VaR modelling and forecasting; see, among others, Taylor (1999,<br />
2000, 2008), Kouretas and Zarangas (2005), Bao et al. (2006) and Kuester et<br />
al. (2006). The distinctive feature of our study is the analysis of the predictive<br />
ability of the observable variables that the proxy for liquidity and trading-activity<br />
in day-ahead VaR modelling. To the best of our knowledge, no previous paper<br />
has focused on this important aspect, although some recent studies can be related<br />
to this analysis. Cenesizoglu and Timmerman (2008) analyze the predictability of<br />
the distribution of monthly returns on a set of state variables that are believed to<br />
predict the equity premium, such as valuation and corporate ratios, bond yields,<br />
and di¤erent measures of default and market risk. Some of these variables are<br />
shown to be helpful in predicting di¤erent quantiles, particularly, at the right<br />
tail of the distribution, which may lead to more e¢ cient portfolio selection and<br />
option trading. Adrian and Brunnermeir (2009) model weekly VaR dynamics in<br />
the banking industry using a similar set of state variables aiming to characterize<br />
the determinants of CoVaR dynamics. Our paper contributes to this literature by<br />
providing novel evidence on i) the predictability of the left tail of the conditional<br />
distribution of daily returns on the basis of volume and liquidity measures, and ii)<br />
the suitability of these variables for downside risk forecasting. This study is also<br />
related to the previous studies focused on the analysis of return predictability and<br />
the links between volatility and the main variables related to the trading ‡ow, such<br />
as bid-ask spreads and trading-based measures; see, among others, Clark (1973),<br />
Tauchen and Pitts (1983), Stoll (1989) and Kalimipalli and Warga (2002).<br />
The remaining part of the paper is organized as follows. Section 2 reviews the<br />
basic elements in VaR modelling and brie‡y introduces the quantile regression approach.<br />
Section 3 describes the main features in the data set analyzed and carries<br />
out the empirical analysis. The main analysis focuses on the volume-weighted market<br />
portfolio, using both (quantile) predictive regressions and backtesting analysis<br />
techniques. For the sake of completeness, we also analyze predictability for di¤erent<br />
characteristic portfolios, namely, value-weighted, book-to-market (B/M) and<br />
size-sorted market portfolios. Finally, Section 4 summarizes and concludes.<br />
3<br />
5
2 Modelling and forecasting downside risk: Value<br />
at Risk<br />
Let r t ; t = 1; :::; T , be the daily log-return time-series of a portfolio, and let F t<br />
be the natural …ltration that includes all the available information at time t; such<br />
as any measurable transformation on the past observations of r t as well as any<br />
other observable variable. For simplicity, but no loss of generality, we assume in<br />
the sequel that returns behave as a stationary martingale di¤erence sequence such<br />
that E (r t jF t 1 ) = 0; with bounded moments E jr t j < 1 for some > 2 large<br />
enough. 4 For a probability 2 (0; 1), we de…ne the (1 )% VaR of a …nancial<br />
asset or portfolio as the maximum loss over a horizon of h days which is expected<br />
at the (1 )% con…dence level given F t , i.e., the -quantile of the conditional<br />
loss distribution of the portfolio. Formally, we can denote:<br />
V aR ;t+h = finf x 2 R : Pr (r t;h xjF t ) g (1)<br />
= fQ ;t (h)g<br />
with Q ;t (h) de…ned implicitly, and r t;h = P j=1;h r t+j denoting the h-period return.<br />
In market risk management, h typically ranges from 1 to 10 days and =<br />
f0:01; 0:05g. For instance, the Basel market risk framework stresses the use of the<br />
1% conditional percentile to determine capital adequacy, while traders in a bank<br />
are often constrained by the rule that the 95% VaR of their position should not<br />
exceed a given bound on a daily basis (McNeil et al., 2005). The extant literature<br />
in …nancial econometrics has suggested very di¤erent procedures with which to<br />
model and forecast VaR dynamics. In the following subsection, we discuss the<br />
main characteristics of the quantile regression approach. Appendix A sketches the<br />
main features of several alternative procedures that shall be used together with<br />
quantile regression in the empirical analysis in Section 4.<br />
2.1 Quantile regression: CAViaR models<br />
Following Koenker and Bassett (1978) and Basset and Koenker (1982), the expected<br />
conditional quantile could be written as a F t -measurable linear function<br />
of n explanatory variables, X t = (x 1t ; :::; x nt ) 0 ; and an (n 1) vector of unknown<br />
coe¢ cients that depends on the -quantile, namely, Q ;t (h ) = X 0 t . Particularizing<br />
for a one-day holding period, h = 1, this formulation is equivalent to<br />
assume the quantile regression model<br />
r t+1 = X 0 t + u t+1; ; t = 1; :::; T (2)<br />
4 This condition is not strictly necessary but simpli…es exposition considerably. It comes with<br />
no loss of generality because it is customary to demean (daily) returns previously to ensure that<br />
the resultant series behaves as a martingale di¤erence sequence; see, for instance, Taylor (2008).<br />
In Section 4, we …lter out the predictable pattern in the conditional mean of the di¤erent portfolio<br />
returns by …tting an AR(1) model prior to apply the VaR methods.<br />
4<br />
6
where u t; is an error term satisfying E (u t; jX t 1 ) = 0: As in the standard regression<br />
model, this general speci…cation does not impose any particular restriction on<br />
the distribution of the data.<br />
Model (2) is highly reminiscent of the standard linear regression setting. Whereas<br />
the least-squares analysis attempts to characterize the conditional mean of the distribution<br />
as a function of the regressors, E (r t+1 jX t ) ; the quantile regression allows<br />
us to model directly the dynamics of the conditional -quantile of the distribution.<br />
A well-known particular case arises for = 1=2 (also known as least absolute<br />
deviation regression model), which is intended to provide estimates of the slope<br />
coe¢ cients for the conditional median of the process. In this case, the relevant<br />
coe¢ cients can be estimated consistently by minimizing the sum of the absolute<br />
values of the residuals. More generally, given an arbitrary value of 2 (0; 1) ; the<br />
unknown vector of parameters in (2) can be estimated consistently as:<br />
( TX<br />
b : arg min ju t+1; j I (rt+1<br />
2R n Xt 0 ) +<br />
t=1<br />
)<br />
TX<br />
(1 ) ju t+1; j I (rt+1
that in the Xt<br />
variable. The main purpose of the autoregressive structure is to<br />
ensure that the dynamics of the conditional quantile change smoothly over time.<br />
Since VaR dynamics are highly persistent, the lag of the dependent process could<br />
also be seen as an instrumental variable that proxies for the true latent process.<br />
Similarly, the variable jr t 1 j is a natural proxy for the unobservable volatility of<br />
returns. Since it introduces a source of (stochastic) short-term variation related to<br />
the arrival of news in the market, this process is expected to be a major driver in<br />
any market risk measure. At this point, the similarities between the basic structure<br />
of the CAViaR model under the restriction = 0 (so-called Symmetric Absolute<br />
Value CAViaR model, Restricted-CAViaR henceforth) and the class of GARCH<br />
models widely used to characterize volatility are fully evident.<br />
In addition to the GARCH-type architecture, the VaR dynamics in the CAViaR<br />
setting could be driven by the lagged values of a set of state variables. The existing<br />
literature has not yet discussed which variables should be included in such an<br />
analysis. For daily market data, the most obvious candidates seem to be the<br />
variables that proxy for liquidity and other trade conditions in the market. The<br />
main problem in this regard is that market liquidity is a di¢ cult magnitude to<br />
measure and one which is not easy to relate to a single aspect of the market. The<br />
central strategy we adopt consists of individually analyzing a number of tradebased<br />
and order-based market variables (see Section 3.1 for details) which are<br />
widely accepted to be related to liquidity in order to detect predictability. Note<br />
that, although the results in a parametric modelling may be sensitive to the choice<br />
of a particular proxy of liquidity or another, we should expect a robust picture to<br />
emerge when considering a wide range of proxying variables. The model resulting<br />
from extending the basic Restricted-CAViaR speci…cation with the lags of a single<br />
predictor can be seen as a low-order individual autoregressive distributed lag model<br />
for the conditional quantile. This class of models is known to be particularly useful<br />
in the forecasting analysis; see, for instance, Rapach and Strauss (2009), and allows<br />
us to examine how a model extended with a variety of individual proxies of liquidity<br />
performs relative to the restricted model.<br />
Finally, under suitable regularity conditions and as the sample size is allowed<br />
to grow unbounded, it can be shown (cf. Engle and Manganelli 2004, Thms. 1<br />
and 2) that:<br />
p<br />
T ( b ) ! d N (0; V ) (5)<br />
i.e., b is a p T -consistent estimate of the unknown vector , and the (suitably rescaled)<br />
estimation bias is asymptotically distributed as a normal distribution with<br />
zero mean and …nite covariance matrix V . In order to consistently estimate this<br />
matrix, Engle and Manganelli (2004) propose a robust estimator that combines<br />
kernel density estimation with the heteroskedasticity-consistent covariance matrix<br />
estimator of White (1980); see Engle and Manganelli (2004, Thm. 3) for details.<br />
We shall use this approach to perform formal inference on the signi…cance of the<br />
estimate.<br />
6<br />
8
3 Empirical analysis<br />
3.1 Data<br />
We consider the returns of di¤erent representative market portfolios in order to<br />
characterize the VaR dynamics. In particular, our basic data set comprises continuously<br />
compounded returns from the volume- and value-weighted portfolios in the<br />
US market over the period 01-04-1988 to 12-31-2002, totaling 3,782 daily observations.<br />
These data are available from CRSP and shall be used in our main analysis<br />
in Section 4.2. In addition, we analyze daily log returns on the B/M-sorted and<br />
size-sorted portfolios in Section 4.3. These data are available at Kenneth French’s<br />
website.<br />
Along with daily portfolio log returns, we observe daily data of several potentially<br />
predictive variables throughout the period analyzed and which are constructed<br />
from individual …rm bid-ask spreads and volume data; see Chordia et al.<br />
(2001) for details. These variables can be classi…ed as:<br />
i) Trading-related variables: Trading Volume (V) measured in thousands of shares;<br />
Number of Trades (NT) calculated as the sum of sell and buy trades; Number of<br />
Sell trades (NS); Number of Shares Sold in thousands (NSS) and Traded Volume<br />
in Dollars (TVD).<br />
ii) Liquidity variables: Quoted Spread (QS) measured as the dollar di¤erence<br />
between ask and bid prices; E¤ective Spread (ES) given by the signed di¤erence<br />
between trade price and bid-ask midpoint (MP); Relative Quoted Spread (RQS)<br />
de…ned as the ratio QS/MP, and Relative E¤ective Spread (RES) de…ned as the<br />
ratio ES/MP. 7<br />
[Insert Table 1 around here]<br />
Table 1 displays in Panel A the usual descriptive statistics for the demeaned<br />
time-series of the di¤erent portfolio returns analyzed in the paper, and all the<br />
explanatory variables (in logarithms) used in our analysis in Panel B. Returns exhibit<br />
the characteristic stylized features in daily samples: Excess of kurtosis, mild<br />
degree of skewness and negligible autocorrelation. The most salient feature of the<br />
predictors in the trade-based and the order-based measures is the strong degree of<br />
persistence as measured by the …rst-order autocorrelation coe¢ cient. Returns are<br />
contemporaneously correlated to all the variables analyzed (correlations are not<br />
shown for restricteding space but are available upon request). In particular, returns<br />
are positively correlated with the variables in the volume group (the average<br />
correlations are around 39%) and negatively correlated to liquidity-related variables<br />
(the average correlations are around -25%). As usual, the variables within<br />
each group are strongly correlated among themselves, and largely and negatively<br />
correlated with the variables in the other group. Cross-correlations range from<br />
-79%, for TVD and QS, to -88%, for TVD and QS.<br />
7 In addition to these variables, we considered alternative variables which did not led to qualitative<br />
changes over those reported in the next section. For instance, considering the logarithm<br />
transformation of the volume or the unexpected volume –measured as the residuals from an<br />
AR(1) model–does not seem to have a major change in the out-of-sample ability of the model.<br />
These results are not presented to save space but are available from the authors upon request.<br />
7<br />
9
Tables<br />
Table 1: Descriptive Statistics.<br />
Panel A: Returns<br />
r t Mean Median Max. Min. Var Skew. Kurt. (1)<br />
r t;V ol 0.00 0.01 5.54 -6.70 0.98 -0.20 7.78 -0.06<br />
r t;V alue 0.02 0.06 4.97 -6.11 0.49 -0.35 10.84 -0.07<br />
r 1t , B=M 0.01 0.02 6.65 -7.85 1.22 -0.11 7.29 -0.07<br />
r 2t , B=M 0.00 0.02 4.81 -6.32 0.67 -0.43 8.36 -0.05<br />
r 1t;size 0.00 0.05 6.02 -7.56 0.66 -0.71 11.06 -0.05<br />
r 2t;size 0.00 0.01 5.73 -6.89 1.05 -0.16 7.39 -0.06<br />
Panel B: Predictive variables<br />
Xt Mean Median Max. Min. Var Skew. Kurt. (1)<br />
V 7.39 7.21 9.69 5.52 0.73 0.36 1.86 0.96<br />
NT 6.85 6.69 8.76 5.07 0.65 0.30 1.64 0.98<br />
NS 6.09 5.94 8.02 4.24 0.65 0.29 1.67 0.98<br />
NSS 6.51 6.33 8.80 4.50 0.73 0.35 1.86 0.96<br />
TVD 11.10 11.04 13.00 9.13 0.76 0.17 1.66 0.96<br />
QS -1.89 -1.74 -1.20 -3.40 0.21 -1.64 4.79 0.99<br />
ES -2.29 -2.11 -1.50 -3.80 0.20 -1.53 4.48 0.99<br />
RQS -5.39 -5.39 -4.80 -6.90 0.20 -1.02 3.23 0.99<br />
RES -5.96 -5.76 -5.20 -7.20 0.20 -0.96 3.08 0.99<br />
This table shows in Panel A the descriptive statistics (mean, median, maximum, minimum,<br />
variance, skewness and kurtosis) of the demeaned returns (previously multiplied by<br />
100) for the volume and value weighted market portfolio, B/M and Size sorted portfolios<br />
corresponding to Low30 (r 1t<br />
) and High30 (r 2t<br />
). Panel B shows the descriptive analysis<br />
for the explanatory variables involved in the analysis (in logarithms). The last column<br />
indicates the …rst-order autocorrelation of the variables. The variables included are V<br />
(trading Volume); NT (Number of Trades); NS (Number of Sell trades); NSS (Number of<br />
Shares Sold in thousands); TVD (Traded Volume in Dollars); QS (Quoted Spread); ES<br />
(E¤ective Spread); RQS (Relative Quoted Spread) and RES (Relative E¤ective Spread).<br />
25<br />
10
3.2 Market portfolio analysis<br />
Throughout the following subsections we analyze the ability of the predictive variables<br />
to forecast day-ahead VaR for the returns of the volume-weighted market<br />
portfolio, both through a predictive regression that involves the whole sample<br />
and an out-of-sample predictive analysis in a rolling-window approach. The focus<br />
on day-ahead estimation is consistent with the holding period considered for<br />
internal risk control by most …nancial …rms (Taylor, 2008). Owing to their empirical<br />
relevance in risk management, we are particularly interested in the quantiles<br />
2 f0:01; 0:05g ; but shall also analyze a wider range of probabilities given by<br />
= f0:01; 0:025; 0:05; 0:075g to characterize predictability along the left tail of<br />
the distribution. It should be mentioned that in this section we also considered the<br />
analysis on the returns of the value-weighted market portfolio, which are not presented<br />
for saving space as the qualitative evidence is completely analogous to that<br />
discussed below. The complete analysis is available from authors upon request.<br />
The daily returns of the volume-weighted market portfolio over the total period<br />
analyzed (see Figure 1 below) include di¤erent episodes in terms of activity and<br />
volatility. The beginning of the sample corresponds to the period that followed the<br />
market crash in October 1987. After the extraordinary crash, the volatility of the<br />
market decreased progressively and reverted to normal levels. In 1998, Long-Term<br />
Capital Management failure in the hedge-fund industry led to a massive bailout<br />
by other major banks and investment houses that, in turn, generated an excess<br />
of volatility in the market and which preceded the burst of the dot-com …rms<br />
in 2000. Finally, the data from 2000 onwards show the large excess of volatility<br />
that characterized the market after the burst of the technological bubble. This<br />
particular period, depicted by the grey line in Figure 1, will be analyzed in detail<br />
in the out-of-sample analysis (back test) later on.<br />
[Insert Figure 1 around here]<br />
3.2.1 Predictive regression analysis<br />
We …rst analyze predictability through predictive-like quantile regressions. More<br />
speci…cally, we estimate the unrestricted CAViaR model<br />
V aR ;t = ;0 + ;1 V aR ;t 1 + ;2 jr t 1 j + j log (X t 1 ) j (6)<br />
for any of the conditioning variables X t described in Section 3.1, using the entire<br />
sample, from 01-04-1988 to 12-31-2002, with 3; 782 observations. Our main<br />
interest is to test the suitability of these unrestricted models against a restricted<br />
CAViaR speci…cations based solely on returns. Clearly, a rejection of the null hypothesis<br />
H 0 : = 0 for a speci…c model and quantile provides formal evidence<br />
on the suitability of the posited variable as a predictor of the conditional distribution.<br />
The objective function in the quantile regression minimization problem is<br />
optimized through the Simulated Annealing optimization algorithm (Go¤e, Ferier<br />
and Rogers, 1994), while the asymptotic covariance matrix is estimated using a<br />
robust sandwich-type estimator based on a k-nearest neighbor kernel, as described<br />
8<br />
11
Figures<br />
6<br />
4<br />
2<br />
Market Portfolio Returns<br />
0<br />
2<br />
4<br />
6<br />
Out of Sample Period<br />
8<br />
1988/01/04 1991/10/01 1995/06/27 1999/03/25 2002/12/31<br />
Observations(Data)<br />
Figure 1: Returns of the volume weighted market portfolio.<br />
3.5<br />
3<br />
Restricted<br />
Volume<br />
RQS<br />
ForecastsVaR (5% nominal level)<br />
2.5<br />
2<br />
1.5<br />
1<br />
1998/09/10 1999/09/10 2000/09/10 2001/09/10<br />
Observations (Data)<br />
Figure 2: Forecasts VaR at nominal level 5% of Restricted-CAViaR<br />
model (solid line) and Unrestricted CAViaR model extended with<br />
volume and Relative Quoted Spread (RQS) variable (grey and black<br />
dashed lines respectively).<br />
34<br />
12
in Engle and Manganelli (2004). 8<br />
[Insert Table 2 around here]<br />
Table 2 reports the estimated coe¢ cients and the one-sided robust p-values for<br />
the all the variables for 2 f0:01; 0:05g. Following Engle and Manganelli (2004),<br />
we set the bandwidth-type parameter in the robust estimation of the covariance<br />
matrix to k = 40 and k = 60 for the 1% and 5% quantiles, respectively. Several<br />
features are worth commenting upon. Firstly, the resulting estimates show the<br />
strong degree of persistence in the VaR dynamics as measured by the estimate of<br />
the autoregressive coe¢ cient, b ;1 . The degree of persistence is weaker for =<br />
0:01 because extreme percentiles are more likely to be driven by jumps (outlying<br />
observations) in data generating process of returns, which are expected to exhibit<br />
a more random, irregular behavior. Secondly, and accordingly with the previous<br />
result, the average in‡uence of the volatility process on the VaR estimates, as<br />
measured by the coe¢ cient b ;2 ; becomes more important as the size of the quantile<br />
decreases. These results completely agree with the qualitative evidence discussed,<br />
among others, in Engle and Manganelli (2004). Turning our attention to the<br />
empirical relevance of the di¤erent predictors, all the variables analyzed have a<br />
positive e¤ect on the conditional distribution of returns. Increments in the volumeor<br />
spread-related variables tend to generate larger levels of VaR, as might be<br />
expected from the theoretical considerations and previous evidence discussed in<br />
Section 1. The analysis on the signi…cance of the estimated coe¢ cients o¤ers,<br />
however, mixed results. Whereas the null hypothesis of no predictability is strongly<br />
rejected for any of the variables analyzed for = 0:05, it cannot be rejected at<br />
any of the usual con…dence levels for the 1% percentile.<br />
There are two possible explanations for this …nding. On the one hand, it is possible<br />
that a variable has predictive power at a given probability yet not at another.<br />
Thus, the statistical analysis would indicate a heterogeneous ability in market liquidity<br />
to predict the tail of the return distribution, with extreme percentiles being<br />
harder to foresee. Since the outlying observations that characterize the left tail<br />
of the returns distribution are related to the jump-component of its data generating<br />
process, and since this component is likely to be driven by a non-regular<br />
process, this point seems plausible: outliers and other extreme movements may<br />
be too volatile and irregular to be predicted systematically by covariates other<br />
than market volatility itself. On the other hand, the discrepancies may be the<br />
consequence of statistical distortions, particularly, for small percentiles such as<br />
= 0:01. In fact, there are two major sources of potential biases in this context:<br />
(small-sample) biases stemming from the robust estimation of the covariance matrix<br />
V , and biases resulting from dealing with a target probability which is close<br />
to the boundary.<br />
8 Simulated annealing is a local random-search search algorithm that can accept values that<br />
increase the objective function (rather than lower it) with a probability that decreases as the<br />
number of iterations increases. The main purpose is to prevent the search process from becoming<br />
trapped in local optima, which in addition provides low sensitivity to the choice of the initial<br />
values. To minimize the possibility of getting convergence to local optima, the optimization<br />
process was repeated 1; 000 times over the whole sample.<br />
9<br />
13
Table 2: Inference results from predictive quantile regressions.<br />
X t<br />
= 5% = 1%<br />
^ ;0<br />
^;1 ^;2 ^ <br />
^ ;0<br />
^;1 ^;2 ^ <br />
V -0.04 0.96 0.05 0.01 0.02 0.92 0.11 0.01<br />
(0.00) (0.00) (0.00) (0.00) (0.37) (0.00) (0.06) (0.16)<br />
NT -0.03 0.96 0.05 0.01 0.03 0.92 0.11 0.01<br />
(0.00) (0.00) (0.00) (0.00) (0.25) (0.00) (0.04) (0.13)<br />
NS -0.02 0.96 0.05 0.01 0.03 0.92 0.11 0.01<br />
(0.00) (0.00) (0.00) (0.00) (0.24) (0.00) (0.04) (0.16)<br />
NSS -0.03 0.96 0.05 0.01 0.03 0.92 0.11 0.01<br />
(0.00) (0.00) (0.00) (0.00) (0.27) (0.00) (0.06) (0.19)<br />
TVD -0.06 0.96 0.05 0.01 -0.01 0.92 0.12 0.01<br />
(0.00) (0.00) (0.00) (0.00) (0.41) (0.00) (0.05) (0.12)<br />
QS -0.01 0.97 0.05 0.01 0.07 0.92 0.14 0.01<br />
(0.01) (0.00) (0.00) (0.00) (0.12) (0.00) (0.08) (0.32)<br />
ES -0.00 0.97 0.05 0.00 0.08 0.91 0.15 0.01<br />
(0.40) (0.00) (0.00) (0.01) (0.14) (0.00) (0.11) (0.31)<br />
RQS -0.04 0.97 0.05 0.01 0.04 0.92 0.14 0.01<br />
(0.01) (0.00) (0.00) (0.00) (0.32) (0.00) (0.08) (0.36)<br />
RES -0.08 0.96 0.05 0.02 -0.03 0.89 0.15 0.03<br />
(0.05) (0.00) (0.00) (0.05) (0.34) (0.00) (0.13) (0.18)<br />
This table shows the estimated parameters and robust p-values (in brackets) for the<br />
entire sample and the quantile regression model (4),<br />
V aR ;t = ;0 + ;1 V aR t 1 + ;2 jr t 1 j + j log(X t 1)j);<br />
given = 0:05 and = 0:01: The …rst column shows the volume-related and liquidity<br />
variables Xt<br />
analyzed.<br />
26<br />
14
The theoretical arguments that support consistency in the robust estimation of<br />
V hold asymptotically. As with other robust estimates of the covariance matrix,<br />
the correct choice of the bandwidth-type parameter in the kernel-type estimation<br />
involved plays a critical role, particularly, in …nite samples. Furthermore, when<br />
is close to the boundary limits, the asymptotic variance increases because of<br />
the small density of the observations in that region. The conjunction of these<br />
factors may cause signi…cant biases in the estimates of the covariance matrix and<br />
meaningful distortions in the subsequent hypotheses testing. In order to analyze<br />
the sensitivity of the results to the estimation of V , we estimated the CAViaR<br />
models for 2 with the bandwidth-type parameter taking values over a wide<br />
range, k 2 f10; 30; 50; 70; 90g. Table 3 summarizes the main results from this<br />
analysis, showing the estimates of the and ;2 coe¢ cients and their robust<br />
p-values given k: For the remaining parameters, there was no qualitative changes<br />
and, therefore, we do not present results in order to save space, although these are<br />
available upon request. Table 3 reveals several interesting …ndings. For relatively<br />
large values of in the tail of the distribution, the empirical evidence strongly<br />
supports predictability for all the variables, independently of the value of k; which<br />
allows us to conclude that predictable relationships do exist. Nevertheless, as <br />
decreases towards the zero limit, the statistical evidence of predictability weakens<br />
and eventually vanishes. Overall, the evidence of predictability seems to be more<br />
evident for the variables in the volume-related group, in which the null H 0 : = 0<br />
is rejected for most variables, and (marginally) rejected even for = 0:01. The<br />
overall analysis also reveals that the main statistical conclusions in this analysis do<br />
not seem to be particularly sensitive to the choice of the bandwidth-type parameter<br />
k.<br />
[Insert Table 3 around here]<br />
The second potential source of biases in this context is related to the existence<br />
of statistical distortions in the quantile regression methodology when dealing with<br />
quantiles close to the boundaries. This well-known problem originates from the<br />
fact that conventional large sample theory does not apply su¢ ciently far in the<br />
tails, a behavior which worsens in …nite samples and which may lead to unsound<br />
inference; see, for instance, Chernozhukov (2005) for a detailed discussion. As<br />
a result, inference may be misleading. Note, for instance, that the signi…cance<br />
analysis of b ;2 in Table 3 leads to puzzling conclusions. Whereas the estimates<br />
of the coe¢ cient related to market volatility largely increases as is smaller,<br />
standard errors increase reducing the degree of statistical signi…cance and causing<br />
H 0 : ;2 = 0 not to be rejected for some nominal levels. This suggests that VaR<br />
dynamics may not be driven by volatility, precisely in the context values in which<br />
total market volatility is the most likely driver of process. This counterintuitive<br />
feature seems a statistical artifact rather than a true rejection of signi…cance.<br />
Owing to the statistical di¢ culties in dealing with extremes percentiles, empirical<br />
studies often avoid the analysis for values of close to boundaries; see, for<br />
instance, Cenesizoglu and Timmerman (2008). In our context, however, we have<br />
alternative methods with which to study the forecasting ability of the predictive<br />
variables even for such percentiles. Since the ultimate purpose of VaR modelling is<br />
to generate accurate market risk forecasts, the most important question is whether<br />
10<br />
15
Table 3: Sensitivity analysis of p-values to di¤erent k-bandwidth.<br />
V NT NS NSS TVD QS ES RQS RES jr t 1 j<br />
7.5% ^ 0.01 0.01 0.01 0.01 0.01 0.02 0.00 0.01 0.02 0.06<br />
10 (0.00) (0.00) (0.00) (0.00) (0.24) (0.00) (0.02) (0.00) (0.13) (0.00)<br />
30 (0.00) (0.01) (0.00) (0.00) (0.00) (0.00) (0.04) (0.07) (0.02) (0.00)<br />
k 50 (0.00) (0.01) (0.00) (0.00) (0.00) (0.00) (0.02) (0.06) (0.06) (0.00)<br />
70 (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.03) (0.07) (0.04) (0.00)<br />
90 (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.02) (0.04) (0.02) (0.00)<br />
5.0% ^ 0.01 0.01 0.01 0.01 0.01 0.01 0.00 0.01 0.02 0.05<br />
10 (0.02) (0.00) (0.01) (0.03) (0.00) (0.00) (0.00) (0.00) (0.16) (0.00)<br />
30 (0.00) (0.01) (0.00) (0.00) (0.00) (0.04) (0.00) (0.00) (0.06) (0.00)<br />
k 50 (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.01) (0.00) (0.04) (0.00)<br />
70 (0.00) (0.00) (0.00) (0.01) (0.00) (0.00) (0.02) (0.00) (0.07) (0.00)<br />
90 (0.01) (0.00) (0.01) (0.01) (0.01) (0.01) (0.00) (0.01) (0.05) (0.00)<br />
2.5% ^ 0.02 0.02 0.02 0.03 0.02 0.01 0.00 0.00 0.03 0.07<br />
10 (0.04) (0.14) (0.43) (0.00) (0.00) (0.04) (0.34) (0.39) (0.46) (0.01)<br />
30 (0.03) (0.17) (0.10) (0.04) (0.07) (0.10) (0.29) (0.36) (0.08) (0.11)<br />
k 50 (0.02) (0.08) (0.06) (0.02) (0.04) (0.12) (0.26) (0.35) (0.11) (0.09)<br />
70 (0.03) (0.03) (0.04) (0.02) (0.02) (0.11) (0.27) (0.35) (0.06) (0.07)<br />
90 (0.03) (0.03) (0.03) (0.02) (0.02) (0.07) (0.28) (0.35) (0.04) (0.07)<br />
1.0% ^ 0.01 0.01 0.01 0.01 0.01 0.01 0.01 0.01 0.03 0.14<br />
10 (0.04) (0.02) (0.05) (0.11) (0.03) (0.24) (0.03) (0.30) (0.00) (0.03)<br />
30 (0.15) (0.16) (0.12) (0.15) (0.14) (0.34) (0.33) (0.38) (0.24) (0.05)<br />
k 50 (0.12) (0.12) (0.11) (0.14) (0.11) (0.33) (0.28) (0.38) (0.15) (0.03)<br />
70 (0.14) (0.15) (0.14) (0.16) (0.10) (0.28) (0.24) (0.30) (0.16) (0.03)<br />
90 (0.11) (0.13) (0.11) (0.12) (0.10) (0.28) (0.24) (0.32) (0.15) (0.03)<br />
This table shows the estimated coe¢ cients ^ ;2 and ^ and robust p-values (in brackets<br />
and in bold) of the test for individual signi…cance from model (4) and the entire sample<br />
when the robust asymptotic covariance matrix is estimated with a kernel with values of<br />
k 2 f10; 30; 50; 70; 90g in the covariance matrix estimation process for a larger set<br />
of quantiles. The ^ ;2 estimates (last column) are from model (4) with Xt = V:<br />
27<br />
16
the covariate-extended risk models have real out-of-sample predictability or not.<br />
Therefore, the backtesting analysis provides us with a more appropriate framework<br />
with which to compare the relative performance of the models, and to attempt to<br />
disentangle the predictive ability of the model for any quantile. This issue is<br />
studied carefully in the following subsection.<br />
3.2.2 Out-of-sample analysis: backtesting analysis<br />
In this section we analyze the day-ahead out-of-sample performance of the covariateextended<br />
CAViaR models relative to the restricted Restricted-CAViaR for 2 :<br />
In addition, we also analyze the relative performance of the covariate-extended<br />
model against several established methods that are based solely on returns, such<br />
as the EWMA model (RiskMetrics), the Gaussian GARCH(1,1) model, and a<br />
model that combines GARCH estimation with a block-maxima approach in the<br />
Extreme Value Theory. The main characteristics of these alternative procedures<br />
are sketched in Appendix A; see McNeil et al. (2005) for further details.<br />
Following Engle and Manganelli (2004) and Alexander and Sheedy (2008) we<br />
consider a rolling-window estimation period formed with the most recent 2; 700<br />
observations available at any day to initialize the risk models and generate VaR<br />
forecasts for the probabilities 2 . This allows us to construct a sequence<br />
of N = 1; 000 one-day ahead daily VaR predictions that can be compared with<br />
realized returns. 9 Whereas backtesting is a crucial element in the validation process<br />
of an internal risk model, a de…nitive methodology has yet to be determined. In<br />
the absence of a regulatory setting, the existing literature has adopted certain<br />
statistical standards that we brie‡y discuss in the sequel. Appendix B provides a<br />
more detailed discussion.<br />
The most popular backtesting procedures analyze the statistical properties of<br />
an indicator variable (known as exception or VaR violation) that signals the occurrence<br />
of a realized return exceeding the expected VaR level, i.e., a random<br />
process, here denoted as H ;t , that takes value one if r t < V aR ; and zero otherwise,<br />
t = 1; :::; N. A desirable property is that the realized number of exceptions<br />
should represent approximately a % of the total out-of-sample period. This property<br />
is known in risk management as reliability and suggests the testable condition<br />
H 0;U : E (H ;t ) = : Also, the dynamic properties of H ;t should resemble those of<br />
an independent and identically distributed process, which combined with the reliability<br />
property renders the joint property of perfect conditional coverage, namely,<br />
H 0;C : E (H ;t jF t 1 ) = .<br />
Christo¤ersen (1988) proposed a sequence of likelihood-ratio tests to check<br />
the properties of i) correct unconditional coverage, ii) (…rst-order serial) independence,<br />
and iii) correct conditional coverage in the exception variable given i) and<br />
ii); see Appendix B for details. These tests are widely used in practice and, therefore,<br />
constitute our …rst backtesting approach. In the sequel, we shall refer to<br />
these procedures as LR UC , LR IND and LR CC , respectively. Similarly, Engle and<br />
Manganelli (2004) proposed another exception-based testing procedure to test for<br />
9 The average estimates of the parameter estimates over the out-of-sample period show the<br />
same patterns discussed before and, hence, are not presented in order to save space, although<br />
these results are available upon request.<br />
11<br />
17
the crucial H 0;C hypothesis, the so-called dynamic quantile test (henceforth DQ).<br />
This test is widely believed to have better properties than LR CC because it has<br />
power to detect higher-order dependences as it explodes a richer set of information.<br />
Consequently, we shall use the DQ procedure in our analysis. Finally, Piazza et<br />
al. (2009) have recently proposed an encompassing VaR quantile-regression test<br />
intended to test H 0;C : E (H ;t jF t 1 ) = : The distinctive characteristic is that<br />
the test is not de…ned on the exception variable, but on the time-series of predictions<br />
itself. This procedure, denoted as V QR in the sequel, uses a linear quantile<br />
regression to address if the VaR model is correctly speci…ed. It should be noted<br />
that the test is asymptotically equivalent to DQ, yet, as claimed by the authors,<br />
may have better properties in …nite samples. Inference on the basis of quantile<br />
regressions, on the other hand, may su¤er from statistical problems when dealing<br />
with extreme quantiles in …nite samples, as discussed previously.<br />
[Insert Table 4 around here]<br />
We …rst analyze the backtesting results for the risk models based solely on returns:<br />
VaR-GARCH, VaR-EWMA, EVT-BM (extreme value theory) and Restricted-<br />
CAViaR models. Table 4 reports the main outcomes from this analysis, displaying<br />
the empirical frequency of exceptions b H = P N<br />
t=1 H ;t=N; the test statistics of the<br />
LR UC , LR IND , LR CC ; DQ and V QR testing procedures as well as their respective<br />
p-values. For all these models, the empirical unconditional coverage for the percentiles<br />
0:05 tend to be greater than the nominal level, with b H signi…cantly<br />
departing from in most cases. Consequently, the hypotheses of perfect conditional<br />
and unconditional coverages tend to be rejected. Similar evidence has been<br />
reported previously, for instance, in Taylor (2008) for GARCH and CAViaR-type<br />
models; see also Piazza et al. (2009).<br />
For quantiles < 0:05; the distortions in the unconditional coverage are considerably<br />
reduced for all but the EWMA model and, therefore, the H 0;U hypothesis<br />
tends not to be rejected. The overall evidence for H 0;C , however, is mixed and depends<br />
on the particular test applied. Whereas the conservative LR CC test accepts<br />
the hypothesis of perfect coverage, the DQ and V QR tests, which are considered<br />
as much more powerful testing procedures, tend to reject the null hypothesis of<br />
correct conditional coverage. Among the di¤erent returns-based VaR procedures<br />
analyzed, the EVT method seems to yield the best performance but, overall, none<br />
of these procedures seems able to pass convincingly the backtesting analysis. Remarkably,<br />
the V QR test rejects the correct performance of all the returns-based<br />
models for any of the conditional quantiles analyzed.<br />
[Insert Tables 5 around here]<br />
Next, we turn our attention to the backtesting results from the volume- and<br />
spread-extended CAViaR models. Table 5 displays the main backtesting results.<br />
The most remarkable feature is that, whereas the standard Restricted-CAViaR<br />
speci…cation and other returns-based approaches exhibit large biases, the inclusion<br />
of the proxies for market liquidity and trading activity conditions improves<br />
the out-of-sample results considerably. The estimated VaR dynamics are shifted<br />
12<br />
18
Table 4: Backtesting VaR analysis for benchmark models. Volume weighted<br />
market portfolio.<br />
MODEL Exc. LR UC LR IND LR CC DQ VQR<br />
EWMA 7.5% 8.9% 2.68(0.10) 0.67(0.41) 3.38(0.18) 23.45(0.00) 14.11(0.00)<br />
5.0% 5.5% 0.51(0.47) 0.00(0.99) 0.52(0.77) 11.40(0.07) 23.54(0.00)<br />
2.5% 1.5% 4.78(0.02) 0.46(0.49) 5.21(0.07) 18.48(0.00) 84.25(0.00)<br />
1.0% 0.5% 3.09(0.08) 0.05(0.82) 3.13(0.21) 3.45(0.74) 177.84(0.00)<br />
GARCH(1,1) 7.5% 11.3% 18.22(0.00) 0.29(0.59) 18.59(0.00) 44.21(0.00) 20.85(0.00)<br />
5.0% 7.4% 10.63(0.00) 0.46(0.49) 11.15(0.00) 29.44(0.00) 31.05(0.00)<br />
2.5% 2.8% 0.35(0.55) 0.08(0.78) 0.44(0.80) 22.51(0.00) 60.93(0.00)<br />
1.0% 0.9% 0.10(0.75) 0.16(0.69) 0.27(0.87) 11.40(0.07) 23.54(0.00)<br />
EVT-BM 7.5% 10.4% 10.85(0.00) 0.58(0.45) 11.49(0.00) 18.24(0.00) 14.91(0.00)<br />
5.0% 6.1% 2.36(0.12) 0.51(0.48) 2.89(0.23) 9.31(0.15) 10.10(0.01)<br />
2.5% 2.4% 0.04(0.83) 0.31(0.58) 0.35(0.84) 3.32(0.76) 46.30(0.00)<br />
1.0% 0.5% 3.10(0.08) 0.04(0.84) 3.13(0.21) 3.64(0.72) 74.36(0.00)<br />
SAV-CAViaR 7.5% 10.1% 8.86(0.00) 0.46(0.49) 8.74(0.01) 28.14(0.00) 9.53(0.01)<br />
5.0% 7.4% 10.63(0.00) 0.50(0.48) 11.19(0.00) 28.91(0.00) 9.63(0.01)<br />
2.5% 3.2% 1.85(0.17) 0.00(0.99) 1.86(0.39) 10.86(0.09) 11.77(0.00)<br />
1.0% 1.3% 0.83(0.36) 0.31(0.57) 1.15(0.56) 14.05(0.02) 27.28(0.00)<br />
This table shows the Backtesting analysis for the one-day forecasts of the VaR given the<br />
EMWA, Gaussian GARCH, EVT and the Restricted-CAViaR (SAV-CAViaR) model<br />
using the volume weighted market portfolio. The second column shows the estimated<br />
ratio of empirical exceptions. LR UC , LR IND , LR CC , DQ and VQR denote the values<br />
of the test statistics for unconditional coverage, independence, conditional coverage, DQ<br />
test and VQR test, respectively, (see Appendix B for details), whereas the p-values of<br />
the respective test statistics are exhibit in brackets.<br />
28<br />
19
Table 5: Backtesting VaR analysis for volume and liquidity extended CAViaR<br />
models. Volume weighted market portfolio. See details in Table 4.<br />
X t Exc. LR UC LR IND LR CC DQ VQR<br />
V 7.5% 8.1% 0.51(0.48) 0.06(0.80) 0.43(0.81) 8.03(0.24) 1.72(0.42)<br />
5.0% 5.5% 0.51(0.47) 0.36(0.55) 0.88(0.64) 7.98(0.23) 1.69(0.43)<br />
2.5% 2.2% 0.38(0.53) 0.94(0.33) 1.32(0.51) 6.02(0.42) 5.36(0.07)<br />
1.0% 0.7% 1.02(0.31) 0.08(0.77) 1.09(0.58) 1.38(0.96) 29.78(0.00)<br />
NT 7.5% 8.1% 0.51(0.48) 0.06(0.80) 0.43(0.81) 7.50(0.27) 1.10(0.58)<br />
5.0% 6.0% 1.98(0.16) 0.61(0.43) 2.61(0.27) 10.92(0.09) 3.15(0.21)<br />
2.5% 2.3% 0.17(0.68) 1.03(0.30) 1.20(0.55) 4.02(0.67) 2.13(0.34)<br />
1.0% 0.9% 0.10(0.75) 0.14(0.70) 0.25(0.88) 2.30(0.88) 33.69(0.00)<br />
NS 7.5% 8.2% 0.69(0.41) 0.03(0.85) 0.55(0.76) 7.06(0.31) 1.42(0.49)<br />
5.0% 5.5% 0.51(0.47) 0.36(0.55) 0.88(0.64) 7.63(0.26) 1.56(0.46)<br />
2.5% 2.0% 1.10(0.29) 0.78(0.38) 1.87(0.39) 5.24(0.51) 5.29(0.07)<br />
1.0% 0.6% 1.89(0.17) 0.06(0.81) 1.94(0.38) 1.99(0.92) 32.95(0.00)<br />
NSS 7.5% 8.0% 0.35(0.55) 0.01(0.91) 0.25(0.88) 6.29(0.39) 1.72(0.42)<br />
5.0% 5.6% 0.73(0.39) 0.28(0.59) 1.03(0.60) 9.53(0.16) 3.39(0.18)<br />
2.5% 1.9% 1.61(0.20) 0.70(0.40) 2.29(0.32) 5.63(0.46) 7.44(0.02)<br />
1.0% 0.6% 1.89(0.17) 0.06(0.81) 1.94(0.38) 2.15(0.90) 53.83(0.00)<br />
TVD 7.5% 8.3% 0.89(0.34) 0.01(0.91) 0.71(0.69) 6.96(0.32) 2.50(0.29)<br />
5.0% 6.1% 2.39(0.12) 1.44(0.23) 3.86(0.14) 13.58(0.03) 3.97(0.14)<br />
2.5% 2.8% 0.36(0.55) 0.08(0.78) 0.44(0.80) 4.47(0.61) 2.13(0.34)<br />
1.0% 1.2% 0.38(0.54) 2.45(0.12) 2.83(0.24) 21.14(0.00) 47.69(0.00)<br />
QS 7.5% 8.6% 0.51(0.48) 0.06(0.80) 0.43(0.81) 8.21(0.22) 5.53(0.06)<br />
5.0% 5.3% 0.18(0.67) 0.55(0.46) 0.75(0.69) 10.34(0.11) 5.39(0.07)<br />
2.5% 2.1% 0.69(0.40) 0.86(0.35) 1.54(0.46) 8.57(0.19) 21.89(0.00)<br />
1.0% 1.0% 0.00(1.00) 0.18(0.67) 0.18(0.91) 14.68(0.02) 95.41(0.00)<br />
ES 7.5% 8.0% 0.35(0.55) 0.10(0.75) 0.34(0.84) 9.02(0.17) 5.26(0.07)<br />
5.0% 5.0% 0.00(1.00) 0.92(0.34) 0.92(0.63) 8.89(0.17) 4.56(0.10)<br />
2.5% 2.2% 0.38(0.53) 0.94(0.33) 1.32(0.51) 7.96(0.24) 23.53(0.00)<br />
1.0% 0.9% 0.10(0.75) 0.14(0.70) 0.25(0.88) 13.84(0.03) 37.52(0.00)<br />
RQS 7.5% 8.6% 1.67(0.19) 0.09(0.75) 1.50(0.47) 0.08(0.08) 3.93(0.14)<br />
5.0% 5.5% 0.51(0.47) 0.36(0.55) 0.88(0.64) 12.46(0.05) 2.68(0.26)<br />
2.5% 2.4% 0.04(0.84) 1.13(0.28) 1.17(0.55) 13.43(0.03) 12.78(0.00)<br />
1.0% 0.7% 1.02(0.31) 0.08(0.28) 1.09(0.58) 1.56(0.95) 32.18(0.00)<br />
RES 7.5% 7.9% 0.23(0.63) 0.00(0.97) 0.13(0.93) 9.02(0.17) 5.46(0.07)<br />
5.0% 5.5% 0.51(0.47) 0.36(0.55) 0.88(0.64) 10.94(0.09) 1.44(0.46)<br />
2.5% 2.4% 0.04(0.84) 1.13(0.29) 1.17(0.56) 7.53(0.27) 4.71(0.10)<br />
1.0% 1.0% 0.00(1.00) 0.18(0.67) 0.18(0.91) 12.59(0.05) 38.64(0.00)<br />
29<br />
20
(see, Figure 2 and comments below for a discussion) in the correct direction such<br />
that most of the empirical departures from the theoretical coverage are removed.<br />
The empirical exception rates tend to stabilize around the nominal levels without<br />
generating dependence or clustering in the exceptions. As a result, all covariateextended<br />
CAViaR models are able to amply pass the three backtesting analyses of<br />
Christo¤ersen (1988) at any of the usual con…dence levels. Similarly, the DQ test<br />
tends to largely support the suitability of the extended models, showing sizeable<br />
statistical gains with respect to the restricted case in the vast majority of cases<br />
analyzed. Finally, and in sharp contrast to the results from returns-based methodologies,<br />
the V QR test yields supportive evidence for correct conditional coverage<br />
in the extended quantile regression setting, particularly, for the set of variables in<br />
the volume group and for quantiles larger than 1%: For the 1% quantile, however,<br />
the V QR rejects the null hypothesis of correct coverage. In view of the overall success,<br />
particularly for variables such as Number of Trades, for which the remaining<br />
testing procedures strongly support the correct coverage hypothesis. Under the<br />
V QR metric, therefore, none of the di¤erent models and methodologies used in<br />
our analysis are able to pass the back tests when addressing the 1% percentile. 10<br />
Coupled with the evidence presented in the predictive-regression analysis, the<br />
overall evidence in the out-of-sample results allows us to conclude that the degree<br />
of liquidity and the trading conditions that characterize the market, as proxyed<br />
by the di¤erent variables considered and, particularly, by those related to volume<br />
characteristics, are predictors of the day-ahead conditional distribution of daily<br />
returns and improve the out-of-sample performance of the VaR forecasts. For<br />
extreme quantiles, such as 1%, the statistical evidence in our analysis is less conclusive:<br />
while the testing procedures based on quantile-regression inference do not<br />
support predictability for this percentile, the remaining back tests analyses do.<br />
[Insert Figure 2 around here]<br />
It is interesting to discuss in greater detail the di¤erences between the forecasts<br />
from the restricted and unrestricted CAViaR models. To this end, Figure<br />
2 displays the one-day 95% VaR forecasts from the restricted Restricted-CAViaR<br />
model (solid line) against the unrestricted CAViaR model extended with either<br />
Relative Quoted Spread (RQS) or Volume (black and grey dashed lines respectively).<br />
As shown in Table 4, the actual proportion of VaR exceptions from the<br />
Restricted-CAViaR model over the sample is much higher than the expected 5%,<br />
and consequently the model is biased towards underestimating the actual level of<br />
market risk in our sample. As depicted in Figure 2, the inclusion of the proxies for<br />
market liquidity and trading activity generates an upward shift in the dynamics of<br />
the predicted VaR process and introduces further variability in the forecasts with<br />
respect to the predictions from the restricted model. As a result, the gap between<br />
10 This may be symptomatic of size distortions. Piazza et al. (2009) analyze the small-sample<br />
properties of the test through Monte Carlo simulation. Under experimental conditions, it is<br />
shown that the VQR tends to reject the null model more frequently than expected, particularly,<br />
for small quantiles such as = 0:01. On real data, the true characteristics of the unknown data<br />
generating process may interfere with the asymptotic properties of the test and introduce further<br />
distortions.<br />
13<br />
21
Figures<br />
6<br />
4<br />
2<br />
Market Portfolio Returns<br />
0<br />
2<br />
4<br />
6<br />
Out of Sample Period<br />
8<br />
1988/01/04 1991/10/01 1995/06/27 1999/03/25 2002/12/31<br />
Observations(Data)<br />
Figure 1: Returns of the volume weighted market portfolio.<br />
3.5<br />
3<br />
Restricted<br />
Volume<br />
RQS<br />
ForecastsVaR (5% nominal level)<br />
2.5<br />
2<br />
1.5<br />
1<br />
1998/09/10 1999/09/10 2000/09/10 2001/09/10<br />
Observations (Data)<br />
Figure 2: Forecasts VaR at nominal level 5% of Restricted-CAViaR<br />
model (solid line) and Unrestricted CAViaR model extended with<br />
volume and Relative Quoted Spread (RQS) variable (grey and black<br />
dashed lines respectively).<br />
34<br />
22
the expected and the actual proportion of exceptions is reduced, which improves<br />
the unconditional performance of the model, as discussed previously. Furthermore,<br />
and as is revealed from the back test analysis, this improvement is achieved without<br />
generating clusters of patches of exceptions. In view of overwhelming statistical<br />
support given by all the backtesting procedures and the quantile predictive regression,<br />
we therefore must conclude that the VaR predictions from the risk models<br />
extended with the variables that proxy for liquidity and, particularly, volume, are<br />
able to track the dynamics followed by the true VaR process more closely than a<br />
purely-returns based model.<br />
3.3 B/M and size portfolios<br />
It is interesting to analyze the quantile-predictability of other representative portfolios.<br />
Fama and French (1992) found that B/M and size characteristics seem to<br />
capture most of the cross-section of average stock returns. Also, it is usual that<br />
investors consider risk pro…les based on growth/value and size to make …nancial<br />
decisions. Hence, it deserves interest to analyze predictability at di¤erent quantiles<br />
for B/M-sorted (Low30 and High30) and Size-sorted (Low30 and High30)<br />
portfolios. The results based on predictive quantile regressions are similar to those<br />
discussed previously. We therefore discuss the main results for the out-of-sample<br />
analysis, since some meaningful di¤erences worthy of comment arise in this case.<br />
[Insert Tables 6.1 and 6.2 around here]<br />
Tables 6.1 and 6.2 report the main outcomes from the backtesting analysis of<br />
the CAViaR models extended with trade- and order-related variables, respectively,<br />
for the Low30 and High30 B/M portfolios given 2 11 . There are meaningful<br />
di¤erences across these portfolios. The LR UC and LR CC tests tend to indicate<br />
a correct performance for all the covariate-extended models for both portfolios.<br />
However, the DQ and, particularly, the V QR test, show a much more conservative<br />
picture about the overall predictability of the Low30 portfolio, particularly, for<br />
spread-related variables. In contrast, the overall evidence of predictability for the<br />
High30 B/M portfolio is much stronger. All the back tests, including V QR, suggest<br />
that the spread-related variables (particularly, e¤ective and relative e¤ective<br />
spread) tend to predict correctly the distribution of the return at the quantiles<br />
analyzed. Therefore, the extent of empirical predictability varies attending to<br />
growth/value pro…les of the portfolios involved, which in turn favours certain variables<br />
over others. As in the case of the market portfolio, the V QR test rejects<br />
the correct performance of all the returns-based risk models for all the quantiles<br />
2 involved, both in the Low30 and High30 portfolios (results not reported<br />
for saving space). Therefore, the overall evidence suggests that models including<br />
state variables that capture di¤erent aspects of the trading process outperform<br />
models based on returns exclusively.<br />
[Insert Tables 7.1 and 7.2 around here]<br />
11 The results of the independence test LR IND are not presented in this section in order to<br />
save space, although these results are available upon request.<br />
14<br />
23
Table 6.1: Backtesting VaR analysis for volume extended CAViaR models. B/M portfolio (Low 30 and High30). See details in<br />
Table 4.<br />
B/M (Low30)<br />
X t Exc. LRUC LRCC DQ VQR<br />
V 7.5% 8.7% 1.98(0.15) 1.72(0.42) 11.92(0.06) 2.09(0.35)<br />
5.0% 5.5% 0.51(0.47) 0.52(0.77) 8.05(0.23) 2.11(0.35)<br />
2.5% 2.3% 0.16(0.68) 0.56(0.75) 5.44(0.49) 6.09(0.05)<br />
1.0% 1.0% 0.00(1.00) 3.17(0.20) 14.32(0.03) 15.30(0.00)<br />
NT 7.5% 8.8% 2.31(0.12) 2.06(0.35) 11.92(0.06) 2.09(0.35)<br />
5.0% 5.9% 1.61(0.20) 1.69(0.42) 15.57(0.02) 2.38(0.30)<br />
2.5% 2.5% 0.04(0.83) 2.47(0.28) 7.98(0.24) 3.71(0.16)<br />
1.0% 1.0% 0.00(1.00) 3.17(0.20) 12.86(0.05) 23.05(0.00)<br />
NS 7.5% 8.6% 1.67(0.19) 1.41(0.49) 14.61(0.02) 2.06(0.36)<br />
5.0% 5.4% 0.33(0.56) 0.34(0.84) 8.01(0.24) 1.23(0.54)<br />
2.5% 2.4% 0.04(0.83) 2.47(0.28) 8.27(0.22) 2.30(0.30)<br />
1.0% 0.7% 1.01(0.31) 5.73(0.05) 19.33(0.00) 45.13(0.00)<br />
NSS 7.5% 8.4% 1.12(0.28) 0.91(0.63) 11.36(0.08) 2.51(0.29)<br />
5.0% 5.6% 0.73(0.39) 0.74(0.68) 10.82(0.09) 2.20(0.33)<br />
2.5% 2.2% 0.38(0.53) 0.87(0.64) 4.74(0.58) 4.30(0.12)<br />
1.0% 0.9% 0.10(0.74) 3.72(0.05) 15.46(0.02) 14.45(0.00)<br />
TVD 7.5% 8.9% 2.67(0.10) 2.43(0.29) 14.04(0.03) 3.40(0.18)<br />
5.0% 6.4% 3.80(0.05) 4.08(0.12) 15.60(0.02) 3.20(0.20)<br />
2.5% 2.7% 0.16(0.68) 1.88(0.38) 6.42(0.38) 5.46(0.07)<br />
1.0% 1.4% 1.43(0.23) 3.31(0.19) 11.66(0.07) 26.72(0.00)<br />
B/M (High30)<br />
Exc. LRUC LRCC DQ VQR<br />
8.8% 2.31(0.12) 2.86(0.23) 4.36(0.63) 3.37(0.19)<br />
6.9% 6.83(0.01) 8.02(0.01) 12.09(0.06) 12.83(0.00)<br />
3.6% 4.37(0.04) 4.46(0.10) 14.77(0.02) 9.87(0.01)<br />
1.3% 0.83(0.36) 1.17(0.55) 9.05(0.17) 4.51(0.10)<br />
8.9% 2.68(0.10) 3.07(0.21) 4.10(0.66) 4.99(0.08)<br />
6.8% 6.17(0.01) 7.51(0.02) 11.96(0.06) 13.26(0.00)<br />
3.1% 1.38(0.24) 1.39(0.49) 7.77(0.25) 1.82(0.40)<br />
1.3% 0.83(0.36) 2.83(0.24) 19.17(0.00) 0.38(0.83)<br />
8.8% 2.32(0.13) 2.86(0.23) 4.16(0.65) 3.85(0.15)<br />
6.8% 6.16(0.01) 7.51(0.02) 12.08(0.06) 12.74(0.00)<br />
3.2% 1.85(0.17) 1.86(0.39) 7.81(0.25) 11.13(0.00)<br />
1.2% 0.38(0.54) 2.67(0.26) 18.49(0.01) 0.38(0.83)<br />
8.9% 2.67(0.10) 3.07(0.21) 4.12(0.66) 3.73(0.15)<br />
6.7% 5.52(0.02) 6.14(0.04) 8.76(0.19) 15.51(0.00)<br />
3.2% 1.84(0.19) 1.86(0.39) 7.70(0.26) 10.47(0.01)<br />
1.5% 2.18(0.13) 2.62(0.26) 17.72(0.00) 15.79(0.00)<br />
9.0% 3.06(0.08) 3.31(0.19) 4.21(0.65) 4.09(0.13)<br />
6.8% 6.16(0.01) 7.51(0.02) 12.33(0.05) 11.12(0.00)<br />
3.4% 2.99(0.08) 3.61(0.16) 10.19(0.12) 6.80(0.03)<br />
1.0% 0.00(1.00) 0.18(0.91) 0.43(0.99) 0.06(0.97)<br />
30<br />
24
Table 6.2: Backtesting VaR analysis for spread extended CAViaR models. B/M portfolio (Low 30 and High30). See details in<br />
Table 4.<br />
B/M (Low30)<br />
X t Exc. LRUC LRCC DQ VQR<br />
QS 7.5% 8.9% 2.67(0.10) 2.43(0.29) 14.19(0.03) 4.89(0.09)<br />
5.0% 5.5% 0.51(0.47) 0.52(0.77) 16.95(0.01) 12.49(0.00)<br />
2.5% 2.1% 0.69(0.40) 1.58(0.45) 12.10(0.06) 80.92(0.00)<br />
1.0% 0.7% 1.01(0.31) 1.09(0.57) 3.29(0.77) 41.48(0.00)<br />
ES 7.5% 8.4% 1.12(0.28) 0.91(0.63) 15.74(0.02) 5.34(0.07)<br />
5.0% 5.4% 0.32(0.56) 0.34(0.84) 0.05(0.05) 10.63(0.00)<br />
2.5% 2.5% 0.00(1.00) 0.23(0.88) 5.55(0.05) 29.39(0.00)<br />
1.0% 0.9% 0.10(0.74) 0.24(0.88) 2.38(0.88) 22.26(0.00)<br />
RQS 7.5% 9.1% 3.47(0.06) 3.28(0.19) 15.45(0.01) 5.31(0.07)<br />
5.0% 5.3% 0.18(0.66) 0.74(0.68) 14.77(0.02) 22.93(0.00)<br />
2.5% 2.5% 0.00(1.00) 0.23(0.88) 9.81(0.13) 55.98(0.00)<br />
1.0% 1.0% 0.00(1.00) 0.18(0.91) 2.23(0.89) 12.81(0.00)<br />
RES 7.5% 8.4% 1.12(0.28) 0.91(0.63) 10.99(0.08) 2.91(0.23)<br />
5.0% 5.6% 0.73(0.39) 1.02(0.59) 14.19(0.02) 22.63(0.00)<br />
2.5% 2.5% 0.00(1.00) 0.23(0.88) 9.20(0.16) 14.33(0.00)<br />
1.0% 1.4% 1.43(0.23) 3.31(0.19) 18.04(0.01) 45.92(0.00)<br />
B/M (High30)<br />
Exc. LRUC LRCC DQ VQR<br />
8.6% 1.67(0.19) 1.89(0.38) 2.69(0.84) 3.54(0.17)<br />
6.1% 2.38(0.12) 2.91(0.23) 7.51(0.27) 4.89(0.09)<br />
2.8% 0.35(0.55) 0.43(0.80) 17.05(0.01) 1.11(0.57)<br />
0.8% 0.43(0.51) 0.55(0.75) 12.28(0.06) 0.14(0.93)<br />
8.2% 0.68(0.40) 1.49(0.47) 2.83(0.83) 2.56(0.28)<br />
5.7% 0.98(0.32) 1.99(0.36) 6.58(0.36) 3.92(0.14)<br />
2.4% 0.04(0.83) 0.34(0.83) 4.76(0.57) 2.62(0.27)<br />
0.9% 0.10(0.74) 0.26(0.87) 20.00(0.00) 0.18(0.91)<br />
8.9% 2.67(0.10) 3.07(0.21) 4.40(0.62) 5.45(0.07)<br />
6.3% 3.29(0.06) 3.65(0.16) 7.16(0.31) 7.67(0.02)<br />
2.9% 0.62(0.42) 0.67(0.71) 16.10(0.01) 3.41(0.18)<br />
1.0% 0.00(1.00) 0.18(0.91) 21.14(0.00) 6.25(0.04)<br />
8.8% 2.31(0.12) 2.87(0.23) 4.09(0.66) 3.67(0.16)<br />
6.3% 3.29(0.06) 4.44(0.11) 7.06(0.32) 8.77(0.01)<br />
2.4% 0.04(0.83) 0.34(0.84) 11.77(0.07) 4.37(0.11)<br />
1.3% 0.83(0.36) 1.15(0.56) 19.21(0.00) 2.53(0.28)<br />
31<br />
25
Table 7.1: Backtesting VaR analysis for volume extended CAViaR models. Size portfolio (Low30 and High30). See details in<br />
Table 4.<br />
SIZE (Low30)<br />
X t Exc. LRUC LRCC DQ VQR<br />
V 7.5% 7.2% 0.13(0.71) 0.29(0.86) 3.77(0.71) 1.41(0.49)<br />
5.0% 3.8% 3.29(0.06) 3.50(0.17) 8.75(0.19) 1.90(0.39)<br />
2.5% 2.3% 0.16(0.68) 1.24(0.53) 2.00(0.92) 1.90(0.39)<br />
1.0% 1.4% 1.43(0.23) 1.84(0.39) 2.84(0.83) 6.13(0.05)<br />
NT 7.5% 7.4% 0.50(0.47) 0.51(0.77) 7.10(7.11) 2.65(0.27)<br />
5.0% 3.9% 2.74(0.09) 2.90(0.23) 7.79(0.25) 2.75(0.25)<br />
2.5% 2.20% 0.38(0.53) 1.36(0.50) 2.25(0.89) 6.01(0.05)<br />
1.0% 1.2% 0.37(0.53) 0.67(0.71) 1.19(0.98) 3.99(0.14)<br />
NS 7.5% 7.3% 0.05(0.80) 0.17(0.91) 4.25(0.64) 4.72(0.09)<br />
5.0% 3.8% 3.29(0.06) 3.50(0.17) 8.72(0.19) 2.80(0.25)<br />
2.5% 2.2% 0.38(0.53) 1.36(0.50) 2.26(0.90) 2.80(0.25)<br />
1.0% 1.2% 0.37(0.53) 0.67(0.71) 1.16(0.98) 7.95(0.02)<br />
NSS 7.5% 7.3% 0.05(0.80) 0.17(0.91) 3.36(0.76) 5.89(0.05)<br />
5.0% 3.8% 3.29(0.06) 0.62(0.17) 8.75(0.19) 3.26(0.20)<br />
2.5% 2.1% 0.69(0.40) 1.58(0.45) 2.68(0.85) 6.16(0.05)<br />
1.0% 1.3% 1.17(0.55) 1.17(0.55) 1.93(0.92) 1.92(0.38)<br />
TVD 7.5% 7.5% 0.00(1.00) 0.04(0.97) 5.15(0.52) 5.10(0.08)<br />
5.0% 3.6% 4.55(0.03) 4.92(0.08) 9.62(0.14) 2.29(0.32)<br />
2.5% 2.3% 0.16(0.68) 1.24(0.53) 2.00(0.92) 7.17(0.03)<br />
1.0% 1.2% 0.37(0.53) 0.67(0.71) 1.16(0.98) 1.01(0.60)<br />
SIZE (High30)<br />
Exc. LRUC LRCC DQ VQR<br />
8.7% 1.98(0.15) 1.72(0.42) 9.98(0.13) 2.17(0.34)<br />
5.9% 1.61(0.20) 2.09(0.35) 9.99(0.12) 2.50(0.29)<br />
2.6% 0.04(0.84) 1.98(0.37) 3.94(0.68) 1.28(0.53)<br />
0.7% 1.01(0.31) 1.09(0.57) 2.80(0.83) 11.96(0.00)<br />
8.7% 1.98(0.15) 1.72(0.42) 9.54(0.15) 1.56(0.46)<br />
6.0% 1.98(0.15) 1.71(0.42) 8.53(0.20) 2.56(0.28)<br />
2.6% 0.04(0.84) 1.98(0.37) 4.76(0.57) 0.78(0.68)<br />
0.7% 1.01(0.31) 1.09(0.57) 2.71(0.84) 14.94(0.00)<br />
8.8% 2.31(0.12) 2.06(0.35) 9.59(0.14) 1.90(0.39)<br />
5.5% 0.51(0.47) 1.73(0.42) 8.99(0.17) 2.50(0.29)<br />
2.4% 0.04(0.83) 0.34(0.83) 2.66(0.85) 2.03(0.36)<br />
0.6% 1.88(0.16) 1.93(0.37) 2.48(0.87) 13.09(0.00)<br />
8.6% 1.67(0.19) 1.41(0.49) 9.31(0.16) 1.32(0.52)<br />
5.8% 1.28(0.25) 1.92(0.38) 9.94(0.13) 1.74(0.42)<br />
2.1% 0.69(0.40) 1.54(0.46) 3.67(0.72) 1.15(0.57)<br />
0.6% 1.88(0.16) 1.93(0.37) 2.63(0.85) 15.20(0.00)<br />
9.2% 3.90(0.04) 3.57(0.16) 12.04(0.06) 3.37(0.19)<br />
6.3% 3.29(0.06) 4.04(0.13) 12.25(0.06) 3.56(0.17)<br />
3.3% 2.38(0.12) 3.12(0.20) 6.53(0.37) 2.22(0.33)<br />
1.0% 0.00(1.00) 3.17(0.20) 12.47(0.05) 10.82(0.00)<br />
32<br />
26
Table 7.2: Backtesting VaR analysis for spreads extended CAViaR models. Size portfolio (Low30 and High30). See details in<br />
Table 4.<br />
SIZE (Low30)<br />
X t Exc. LRUC LRCC DQ VQR<br />
QS 7.5% 7.6% 0.01(0.90) 0.03(0.98) 8.14(0.23) 3.31(0.19)<br />
5.0% 4.1% 1.81(0.17) 1.85(0.39) 10.36(0.11) 8.05(0.02)<br />
2.5% 2.4% 0.04(0.83) 1.22(0.54) 5.43(0.49) 7.06(0.03)<br />
1.0% 1.5% 2.18(0.13) 2.65(0.26) 4.18(0.65) 3.01(0.22)<br />
ES 7.5% 7.2% 0.13(0.71) 0.29(0.86) 9.51(0.15) 3.37(0.19)<br />
5.0% 3.8% 3.29(0.06) 3.47(0.17) 9.61(0.14) 3.34(0.19)<br />
2.5% 2.3% 0.16(0.68) 1.24(0.53) 5.81(0.44) 9.02(0.01)<br />
1.0% 1.5% 2.18(0.13) 2.65(0.26) 4.36(0.63) 2.14(0.34)<br />
RQS 7.5% 7.6% 0.01(0.90) 0.34(0.84) 10.27(0.11) 2.58(0.27)<br />
5.0% 4.1% 1.81(0.17) 1.85(0.39) 9.45(0.15) 11.62(0.00)<br />
2.5% 2.3% 0.16(0.68) 1.24(0.53) 2.03(0.92) 4.21(0.12)<br />
1.0% 1.4% 1.43(0.23) 1.84(0.39) 7.30(0.29) 2.62(0.27)<br />
RES 7.5% 7.3% 0.05(0.80) 0.17(0.91) 7.55(0.27) 3.46(0.18)<br />
5.0% 3.8% 3.29(0.06) 3.47(0.17) 8.84(0.18) 9.69(0.01)<br />
2.5% 2.1% 0.69(0.40) 1.58(0.45) 2.66(0.85) 15.37(0.00)<br />
1.0% 1.4% 1.43(0.23) 1.84(0.39) 2.89(0.82) 2.21(0.33)<br />
SIZE (High30)<br />
Exc. LRUC LRCC DQ VQR<br />
8.4% 1.12(0.28) 0.91(0.63) 13.92(0.03) 6.45(0.04)<br />
5.7% 0.98(0.32) 3.12(0.21) 15.12(0.02) 4.52(0.10)<br />
2.1% 0.69(0.40) 1.54(0.46) 4.91(0.56) 7.87(0.02)<br />
0.6% 1.88(0.16) 1.95(0.37) 3.47(0.75) 76.58(0.00)<br />
8.2% 0.68(0.40) 0.55(0.75) 12.67(0.05) 6.45(0.04)<br />
5.4% 0.32(0.56) 1.77(0.41) 12.76(0.05) 4.12(0.13)<br />
2.1% 0.69(0.40) 1.29(0.52) 4.22(0.65) 10.49(0.01)<br />
0.7% 0.00(1.00) 1.10(0.57) 4.90(0.56) 12.61(0.00)<br />
8.5% 1.38(0.23) 1.14(0.56) 11.38(0.08) 2.58(0.27)<br />
5.8% 1.28(0.25) 1.92(0.38) 17.27(0.01) 4.23(0.12)<br />
2.5% 0.00(1.00) 1.23(0.54) 4.13(0.66) 2.28(0.32)<br />
0.8% 0.43(0.51) 0.54(0.76) 1.86(0.93) 13.29(0.00)<br />
6.9% 0.53(0.46) 0.90(0.63) 11.31(0.08) 2.36(0.31)<br />
5.7% 0.98(0.32) 0.98(0.60) 14.84(0.02) 8.76(0.01)<br />
2.2% 0.38(0.38) 0.87(0.64) 5.28(0.51) 3.52(0.17)<br />
0.5% 3.09(0.07) 3.12(0.20) 2.97(0.81) 70.59(0.00)<br />
33<br />
27
Finally, Tables 7.1 and 7.2 show the main outcomes for the backtesting analysis<br />
on the Low30 and High30 Size-sorted portfolios given the variables in the volume<br />
and liquidity groups, respectively. The results in this analysis are, in general, more<br />
similar to those discussed for the volume-weighted market portfolio. The backtesting<br />
analysis reveals a great performance of both volume- and liquidity-extended<br />
risk models to forecast the VaR of the Low30 portfolio at any of the quantiles. The<br />
best results are observed for volume-related variable, for which all the back tests,<br />
including the V QR, tend to accept the suitability of the model. For the High30<br />
portfolio, however, the results are more conservative and in line with those already<br />
discussed for the volume-weighted portfolio: volume and liquidity proxies seem<br />
to make a good job for most quantiles under the di¤erent tests, with the overall<br />
evidence being more ambiguous for the 1% percentile in the VQR test. This suggests<br />
that the conditional distribution of small capitalization stocks is predictable,<br />
particularly, when including volume-related variables. It should be noted that the<br />
overall performance of returns-based procedure (not presented) seems better for<br />
small capitalization assets, with the extreme value theory models yielding the best<br />
performance throughout. Nevertheless, the performance is inferior to the extended<br />
models and, as in the previous cases, none of the returns-based models are able to<br />
pass the V QR test.<br />
4 Concluding remarks<br />
In this paper, we have analyzed the predictability of the tail of the conditional<br />
distribution of market returns using di¤erent variables that are widely related to<br />
trading activity and market liquidity. Our methodological approach has mainly<br />
built on the quantile regression methodology and, particularly, the CAViaR quantile<br />
regression setting proposed in Engle and Manganelli (2004). This strategy<br />
allows us to study in a simple and direct way the forecasting ability of a number<br />
of covariate-extended models in relation to a restricted model that relies solely on<br />
returns, using both predictive regressions and backtesting analysis.<br />
The extent of statistical evidence supporting predictability may vary depending<br />
mainly on the size of the target quantile. However, the overall evidence strongly<br />
suggests the existence of predictability of the conditional distribution on the basis<br />
of trading activity and liquidity variables. This evidence is particularly strong<br />
for the set of volume-related variables in the market portfolio and, more generally,<br />
is fairly robust against the consideration of di¤erent representative market<br />
portfolios, di¤erent proxy variables of liquidity and trading activity, and di¤erent<br />
testing procedures. Consequently, the main conclusion from this paper is that the<br />
day-ahead conditional distribution of returns is predictable when using observable<br />
market information which is not necessarily limited to returns and, therefore, such<br />
information can be used in downside risk modelling.<br />
Finally, although our analysis has focused on the analysis of quantiles and,<br />
therefore, allows us to obtain direct conclusions for the VaR methodology, the<br />
main conclusions from our analysis may be extrapolated to other quantile-based<br />
downside risk measures, among which expected shortfall (conditional VaR) is the<br />
most representative. If quantiles are predictable, trivial transformations such as<br />
15<br />
28
the mean of quantiles should be predictable as well. The formal analysis of this<br />
interesting issue is left for future research.<br />
16<br />
29
Appendix A: VaR models<br />
In Section 4 we compare the performance of CAViaR model with other standard<br />
procedures to compute VaR. These include the EWMA, GARCH and Extreme<br />
Value Theory methods. The common setting in these parametric models assumes<br />
that (conditionally demeaned) returns obey dynamics given by<br />
r t = t t ; t jF t 1 iidN (0; 1) (A.1)<br />
where t denotes the conditional volatility of the process. We brie‡y discuss the<br />
main settings of these approaches in the sequel.<br />
A. VaR EWMA<br />
RiskMetrics popularized the Exponential Weighting Moving Average (EWMA)<br />
scheme as an easy way to model the volatility process. The latent volatility dynamics<br />
are assumed to obey the recursive dynamics:<br />
b 2 t = b 2 t 1 + (1 )r 2 t 1; t = 1; :::; T (A.2)<br />
with the initial condition b 2 0 = r0 2 = E (rt 2 ) : The smoothing parameter 0 1<br />
can be estimated, although RiskMetrics advises the setting = 0:95 for data<br />
recorded on a daily basis. Then, the one-day ahead forecast given F T is simply<br />
given by b 2 T +1jT = b 2 T + (1 )rT 2 :<br />
RiskMetrics further assumes the particularly strong assumption that the innovations<br />
t are conditionally normal distributed, from which the one-period ahead<br />
VaR forecast would be given by Z b T +1jT ; with Z denoting the -quantile of<br />
the standard normal distribution. In order to ensure robustness against departures<br />
from normality, we proceed in a slightly di¤erent way. Let b t = r t =b t be<br />
the estimated innovations given the estimates of the EWMA volatility process,<br />
and let Q (b t ) be the unconditional -quantile of the empirical distribution of b t .<br />
Then, a ‘robusti…ed’VaR forecast that does not rely upon distributional assumption<br />
is given by:<br />
V aR ;t+1 (EW MA) = Q (b t )b T +1jT (A.3)<br />
B. VaR GARCH<br />
The simplest GARCH (1,1) model is the most popular approach to model and<br />
forecast market risk in practice due to its impressive performance and statistical<br />
properties (Hansen and Lunde, 2005). The standard GARCH(1,1) model assumes<br />
that daily returns obey dynamics given by:<br />
r t = t t ; t jF t 1 iidN (0; 1) (A.4)<br />
2 t = ! + " 2 t 1 + 2 t 1<br />
with the restrictions ! > 0; ; 0 ensuring that the conditional variance<br />
process is well-de…ned. Although …nancial returns are known to be non-normally<br />
distributed, the Gaussian assumption is particularly convenient because it ensures<br />
parameter consistency under certain regularity conditions even in the absence of<br />
17<br />
30
normality. Parameters can thus be estimated by (quasi) maximum likelihood estimation,<br />
yielding a consistent estimate of the conditional variance process. The<br />
day-ahead forecast is then given by:<br />
b 2 T +1jT = b! + br 2 T + b b 2 T<br />
(A.5)<br />
As in the EWMA model discussed previously, given the GARCH estimates b t<br />
and the resultant standardized innovations, b t = r t =b t ; the ‘robust’one-day VaR-<br />
GARCH forecast is determined as:<br />
V aR ;T +1 (GARCH) = Q (b t )b T +1jT : (A.6)<br />
with Q (b t ) denoting the -quantile of the empirical distribution.<br />
C. VaR Extreme Value Theory<br />
This method can be seen as a parametric re…nement of the previous approaches.<br />
Essentially, the procedure requires the characterization of the tail behavior of the<br />
set of i.i.d. innovations t in the return process. To circumvent the problem that t<br />
is not observable directly, the estimated residuals b t = r t =b t can be used instead,<br />
with b t determined according to some volatility model, such as those discussed<br />
previously. Since GARCH estimates tend to outperform any other procedure, we<br />
estimate the empirical process b t on the basis of the GARCH(1,1) estimates as<br />
discussed above.<br />
The rest of the procedure is described as follows. Given the series b t ; the total<br />
sample period is divided into B = 740 blocks of length l = 5 observations to record<br />
the maximum value of each block (i.e., the maximum loss in the period), say m b ,<br />
b = 1; :::; B; in a time-series process. The Extreme Value Theory suggests …tting<br />
the Generalized Extreme Value distribution (GEV, also known as Fisher–Tippett<br />
distribution) to this series. The GEV arises as the limit distribution of properly<br />
normalized maxima of a sequence of i.i.d. random variables, and is characterized<br />
by the density function<br />
f (z b ; 1 ; 2 ; 3 ) =<br />
1 1=3 1<br />
n<br />
[1 + <br />
3 z b ] exp<br />
2<br />
[1 + 3 z b ]<br />
1= 3<br />
o<br />
(A.7)<br />
if z b > 1; where z b = (m b 1 ) = 2 denotes the standardized variable. The<br />
(unknown) parameters characterize the shape ( 1 ), scale ( 2 ) and location ( 3 ) of<br />
the distribution and can be estimated consistently by di¤erent methods, such as<br />
maximum-likelihood. The importance of this approach is that by inverting this<br />
distribution (with the unknown parameters replaced by their consistent estimates),<br />
we can go from the asymptotic GEV distribution of maxima to the distribution of<br />
the observations themselves and obtain a closed-form expression for the unconditional<br />
VaR of b t given ; namely,<br />
"<br />
b<br />
Q (b t ) = b 2<br />
3 1<br />
b 1<br />
<br />
<br />
log 1<br />
# b1<br />
1<br />
l<br />
(A.8)<br />
Finally, as in the EWMA and GARCH approaches, we generate the one-day ahead<br />
18<br />
31
VaR forecast as<br />
with b T +1jT determined as in (A:5) :<br />
V aR ;T +1 (EV T ) = b T +1jT Q (b t ) (A.9)<br />
19<br />
32
Appendix B: Backtesting analysis<br />
I) Unconditional test, LR UC :<br />
The most basic assumption is that the market risk model provides a correct<br />
unconditional coverage, namely, H 0;U : E [H ;t ] = . The null hypothesis is rejected<br />
for large values of the likelihood-ratio test de…ned as<br />
<br />
LR UC = 2(N N ) log(1<br />
<br />
N <br />
N ) log(1 ) + 2N log N <br />
N<br />
<br />
log 2 (1)<br />
P<br />
t=1;N<br />
(B.1)<br />
where 2 (1) stands for a Chi-squared distribution with one degree of freedom, N <br />
H t; is the number of exceptions, and N is the total number of out-ofsample<br />
observations. Note that N =N = b H is simply the sample mean of H ;t ,<br />
i:e, the sample equivalent of E [H ;t ] :<br />
II) Independence test, LR IND :<br />
If exceptions are serially correlated, the property of reliability conditional coverage<br />
will be defective even if the unconditional coverage is correct, because the<br />
risk of bankruptcy is higher. Christo¤ersen (1998) proposes the analysis of the<br />
…rst-order serial correlation in H ;t through a binary …rst-order Markov chain with<br />
transition probability matrix<br />
=<br />
<br />
1 01 01<br />
; with <br />
1 11 ij = Pr(H ;t = j j H ;t 1 = i); i; j 2 f0; 1g (B.2)<br />
01<br />
The approximate joint likelihood conditional on the …rst observation is<br />
L(; H ;t j H ;1 ) = (1 01 ) n 00<br />
n 01<br />
01 (1 11 ) n 10<br />
n 11<br />
11 ; (B.3)<br />
where n ij represents the number of transitions from state i to state j. The<br />
maximum-likelihood estimators under the alternative hypothesis are b 01 = n 01 = (n 00 + n 01 ) ;<br />
and b 11 = n 11 = (n 10 + n 11 ) : Under the null hypothesis of independence, we have<br />
01 = 11 = 0 ; with 0 = ; from which the conditional binomial joint likelihood<br />
is<br />
L( 0 ; H ;t j H ;1 ) = (1 01 ) n 00+n 10<br />
n 01+n 11<br />
01 : (B.4)<br />
Note that 0 can be estimated as b 0 = N =N. The likelihood ratio test for the<br />
hypothesis of independence is given by<br />
h<br />
i<br />
LR IND = 2 log L(^; H ;t j H 1 ) log L(b 0 ; H ;t j H ;1 ) 2 (1)<br />
(B.5)<br />
III) Conditional test, LR CC :<br />
20<br />
33
Finally, we can study simultaneously whether the VaR violations are independent<br />
and occur with the correct probability, i.e., H 0;C : E [H ;t jF t 1 ] = . Because<br />
b 0 is unconstrained, the test in equation (B.5) does not impose the correct coverage.<br />
Christo¤ersen (1998) devised a joint test for independence and correct<br />
coverage (i.e., correct conditional coverage) by combining the previous tests:<br />
h<br />
i<br />
LR CC = 2 log L(^; H ;T j H 1 ) log L(; H T j H 1 ) 2 (2) (B.6)<br />
This is equivalent to testing if the sequence of H ;t is independent and the probabilities<br />
to observe an exception given the set of information is equal to the nominal<br />
level , namely, 01 = 11 = . Therefore, we can write<br />
LR CC = LR UC + LR IND ;<br />
(B.7)<br />
which provides the suitable test statistic to check whether H ;t exhibits correct<br />
conditional coverage properties. Since the test involves two restrictions, the asymptotic<br />
convergence to a 2 (2) distribution.<br />
IV) Dynamic Quantile test, DQ (Engle and Manganelli, 2004).<br />
The LR CC test does not have the power to detect higher-order dependences<br />
in H ;t . Engle and Manganelli (2004) introduced a test that accounts for a more<br />
general form of dependence in order to test H 0;C : E [H ;t jF t 1 ] = . De…ne the<br />
time series:<br />
Hit t = H ;t (B.8)<br />
and let the matrix of instrumental variables<br />
Z t = [Hit t 1 ; Hit t 2 ; :::; Hit t p ; V aR ;t 1 ; :::; V aR ;t q ] : (B.9)<br />
for some predetermined, …xed lag values p; q 1. Note that H 0;C : E [H ;t jF t 1 ] =<br />
implies the martingale condition E [Hit ;t jF t 1 ] = 0, which in turn could tested<br />
as E [Hit ;t jZ t ] = 0 for suitable values of p and q. Following Engle and Manganelli<br />
(2004), we set p = 4; q = 1 and test the martingale restriction through ordinary<br />
least-squares analysis in the auxiliary regression Hit t = Z t + u t by analyzing the<br />
joint restriction H 0 : = 0 through a standard F test, given by the test statistic<br />
DQ = b 0 Z 0 tZ t<br />
b <br />
(1 )<br />
(B.10)<br />
with b denoting the OLS estimate of : Under the null hypothesis, DQ 2 (p+q) .<br />
V) VaR Quantile Regression test, V QR (Piazza et al., 2009).<br />
Given the VaR forecasts V aR ;t , t = 1; :::; N; Piazza et al. (2009) focus on the<br />
encompassing quantile regression<br />
r t = 0; + 1; V aR ;t + " t; ; t = 1; ::; N<br />
(B.11)<br />
where r t denotes the realized returns over the out-of-sample period. The hypothesis<br />
that V aR ;t is an optimal forecasts of the conditional -quantile of r t can be tested<br />
21<br />
34
through the joint restriction H 0 : 0; = 0; 1; = 1 or, equivalently, H 0 : = 0<br />
with = ( 0 ; 1 1) 0 in the previous regression model.<br />
Let b the quantile regression estimate of : Under standard regularity conditions,<br />
the asymptotic distribution of p <br />
N b is a normal with zero mean<br />
and …nite variance . Under the null, it follows that<br />
V QR = T [ b 0 1 b ] 2 (2)<br />
where the covariance matrix can be estimated by usual methods.<br />
(B.12)<br />
References<br />
[1] Adrian, T. and Brunnermeir, M.K. 2009. “CoVaR.” Working Paper, New<br />
York.<br />
[2] Alexander, C. and Sheedy, E. 2008. “Developing a Stress Testing Framework<br />
Based on Market Risk Models.” Journal of Banking and Finance,<br />
32(10):2220-2236.<br />
[3] Bao, Y., Lee, T. and Salto B. 2006. “Evaluating Predictive Performance of<br />
Value-at-Risk Models in Emerging Markets: A Reality Check.” Journal of<br />
Forecasting, 25(2):101-128.<br />
[4] Bassett, G.W. and Koenker, R. 1982. “An Empirical Quantile Function for<br />
Linear Models with iid Errors.”Journal of the American Statistical Association,<br />
77(378):407-415.<br />
[5] Bollerslev, T. and Melvin, M. 1994. “Bid-Ask Spreads and the Volatility in the<br />
Foreign Exchange Market: An Empirical Analysis.”Journal of International<br />
Economics, 36(3-4):355-372.<br />
[6] Cenesizoglu, T. and Timmerman, A. 2008. “Is the Distribution of Stocks<br />
Returns Predictable.”Working Paper, University of California at San Diego.<br />
[7] Chernozhukov, V. 2005. “Extremal Quantile Regression.”The Annals of Statistics,<br />
33(2):806-839.<br />
[8] Chordia, T., Roll, R. and Subrahmanyam, A. 2001. “Market Liquidity and<br />
Trading Activity.”Journal of Finance, 56(2):501-530.<br />
[9] Christo¤ersen, P.F. 1998. “Evaluating Interval Forecasts.”International Economic<br />
Review, 39(4):841-862.<br />
[10] Clark, P.K. 1973. “A Subordinated Stochastic Process Model with Finite Variance<br />
for Speculative Prices.”Econometrica, 41(1):135-156.<br />
[11] Cochrane, J.H. 2005. Asset Pricing (Revised Edition). Princeton University<br />
Press: Princeton and Oxford.<br />
22<br />
35
[12] Fama, E.F. and French, K.R. 1992. “The Cross-Section of Expected Stock<br />
Returns.”Journal of Finance, 47(8):427-465.<br />
[13] Engle, R. and Manganelli, S. 2004. “CAViaR: Conditional Autoregressive<br />
Value at Risk by Regression Quantiles.” Journal of Business and Economic<br />
Statistics, 22(4):367-381.<br />
[14] Go¤e, W.L., Ferrier, G.D. and Rogers, J. 1994. “Global Optimization of Statistical<br />
Functions with Simulated Annealing.”Journal of Econometrics, 60(1-<br />
2):65-99.<br />
[15] Hansen, P.R. and Lunde, A. 2005. “A Forecast Comparison of Volatility Models:<br />
Does Anything Beat a GARCH (1,1)?”Journal of Applied Econometrics,<br />
20(7):873-889.<br />
[16] Kalimipalli, M. and Warga, A. 2002. “Bid/Ask Spread, Volatility and Volume<br />
in the Corporate Bond Market.”The Journal of Fixed Income, 11(4):31-42.<br />
[17] Koenker, R. and Bassett, G. 1978. “Regression Quantiles.” Econometrica,<br />
46(1):33-50.<br />
[18] Kouretas, G.P. and Zarangas, L. 2005. “Conditional Ine¢ ciency Value at Risk<br />
by Regression Quantiles: Estimating Market Risk for Major Stock Markets.”<br />
Working Paper, Preveza, Greece.<br />
[19] Kuester, K., Mittnik, S. and. Paolella, M.S. 2006. “Value-at-Risk Prediction:<br />
A Comparison of Alternative Strategies.”Journal of Financial Econometrics,<br />
4(1):53-89.<br />
[20] Manganelli, S. and Engle,.R.F. 2004. A Comparison of Value-at-Risk Models<br />
in Finance. Wiley: Chichester.<br />
[21] McNeil, A.J., Frey, R. and Embrechts, P. 2005. Quantitative Risk Management:<br />
Concepts, Techniques, and Tools. Princeton University Press: Princeton.<br />
[22] Piazza, W., Renato, L., Linton, O. and Smith, D. 2009. “Evaluating Valueat-Risk<br />
Models via Quantile Regression.” Working Paper. Forthcoming in<br />
Journal of Business and Economic Statistics.<br />
[23] Rapach, D.E. and Strauss, J.K. 2009. “Di¤erences in Housing Price Forecastability<br />
Across U.S. States.” International Journal of Forecasting, 25(2):351-<br />
372.<br />
[24] Stoll, H.S. 1989. “Inferring the Components of the Bid-Ask Spread: Theory<br />
and Empirical Tests.”Journal of Finance, 44(1):115-134.<br />
[25] Suominen, M. 2001. “Trading Volume and Information Revelation in Stock<br />
Markets.”The Journal of Financial and Quantitative Analysis, 36(4):545-565.<br />
[26] Tauchen, G.E. and Pitts, M. 1983. “The Price Variability-Volume Relationship<br />
on Speculative Markets.”Econometrica, 51(2):485-505.<br />
23<br />
36
[27] Taylor, J.W. 1999. “A Quantile Regression Approach to Estimating the Distribution<br />
of Multiperiod Returns.”Journal of Derivatives, 7(1):64-78.<br />
[28] Taylor, J.W. 2000. “A Quantile Regression Neural Network Approach to Estimating<br />
the Conditional Density of Multiperiod Returns.” Journal of Forecasting,<br />
19(4):299-311.<br />
[29] Taylor, J.W. 2008. “Using Exponentially Weighted Quantile Regression to<br />
Estimate Value at Risk and Expected Shortfall.”Journal of Financial Econometrics,<br />
6(3):382-406.<br />
[30] White, H. 1994. Estimation, Inference and Speci…cation Analysis. Cambridge<br />
University Press: Cambridge.<br />
24<br />
37
ad<br />
serie<br />
<strong>Ivie</strong><br />
Guardia Civil, 22 - Esc. 2, 1º<br />
46020 Valencia - Spain<br />
Phone: +34 963 190 050<br />
Fax: +34 963 190 055<br />
Department of Economics<br />
University of Alicante<br />
Campus San Vicente del Raspeig<br />
03071 Alicante - Spain<br />
Phone: +34 965 903 563<br />
Fax: +34 965 903 898<br />
Website: www.ivie.es<br />
E-mail: publicaciones@ivie.es